Try   HackMD

AI Overview Session Plan

Start with Audience

I usually start with asking the audience to volunteer definitions, thoughts on AI. I ask them to tell me What AI is. I wait a few minutes then usually a few participants will start saying ideas. It's important to remind them that there is no good answer, and they can say something moderately silly. If there's no volunteers or very few, I sometimes prompt the audience (anyone thought of robots?)

I write each thought on my iPad (or a whiteboard) and then I spend time discussing each. Saying why they are good thoughts, and perhaps highlighting ways in which they don't fully capture AI.

Humans vs Machines figure

I then say 'Here is how I usually think about AI' and draw a Venn diagram. On the left, I draw the set of tasks humans can accomplish (explaining that this includes everything from filing your taxes, to solving maths problems or doing the dishes) and on the right, overlapping, I draw the set of tasks machines/computation/computers can accomplish.

I then talk about how there are examples everywhere. If I have time I ask participants to volunteer ideas of things only humans can do, or things only machines can do, etc.

I usually focus on examples within chess:

  • playing chess, in terms of finding the next move in a game, is in the intersection. I mention Kasparov vs DeepBlue, etc
  • actually picking up a chess piece and competently manipulating physical objects is something that is still a human ability. I say that you can program a robot to do it, but if you use a different kind of chessboard, or change the weight of the chess pieces, usually robots would still fail. This may change in a few years though…
  • going through a very large dataset of past chess games and couting the number of times a certain pattern occurred is an example of a task computers are typically more competent at. Humans can of course do it, but it usually takes ages.
  • as an example of a task outside both sets, I mention that we don't currently have a mathematical proof whether an optimal strategy exists in chess (one that would guarantee that white is never beaten by black if white follows the strategy). Such proof exists for trivial games like TicTacToe, other games, like Otello, which is a recent example, but for chess it's beyond us.

One set is growing fast

Then I say that while human competence largely stagnated for a long time, the right-hand set of things that computers can autonomously achieve has expanded very rapidly over the past few generations/decades. I draw radial arrows pointing out from the right-hand set in the Venn diagram illustrating that it is getting inflated.

Then I say that AI is always about what happens at this expanding frontier, especially when it comes to computers gaining abilities that were previously reserved for humans. Thanks to the constant expansion, this is a moving target. Playing chess was once considered a challenging task for AI, but it is no longer surprising that chess can be played by computers. After chess, Go, was considered a difficult benchmark. This is significantly more difficult for computers (as well as for humans) to play. Yet in 2017 or so, this has fallen to computers as well. Most recently, it is the generation and understanding of human language which has been 'claimed' by computers. Subtle languange manipulation such as understaning jokes, etc, were thought to be way beyong ML capabilities just 5 years ago.

Examples of very rapid growth

I then usually try to illustrate the extremely rapid growth of AI capabilities over the last decade. A particularly visual way of doing this is through image generation. For this I have to explain what image generation, as a task, is about, which takes a bit of time. But then a comparison like this one I think is striking:

Progress-in-synthetic-face-generation-due-to-advances-in-self-supervised-generative-AI

The greyscale thing on the left labelled 2014 is the original GAN. When this paper and result came out, everyone was very surprised (now of course I can authentically say this since I was there, but this was really very exciting). Very quickly, year by year, the quality of generations has improved. And then I point out that it this progress is absolutely incredible if you consider the expansion in resolution. The thing on the left is 32 by 32 greyscale pixels (That's described by about a 1000 numbers, each of which has to be set to the right value so the whole thing shows a face). The one on the right is 1000x1000 pixels, and in colour.

I sometimes then show current examples of image and video generation. I start with thispersondoesnotexist.com which is the 2018 tech, where every time you randomly refresh, you get a new generated face. Explain that this person does not exist.

Then I show something like DALL-e, a description-based image generator.

Then I show video generation, something like SORA.

Then I show the demo of Channel1.ai of completely AI generated news anchors, go to the time in the video when it demonstrates talking to completely different languages smoothly. I explain how human speech generation is another breakthrough, and now NotebookLM can generate a 20 minute podcast episode that sounds incredibly real about any topic you ask. ChatGPT advanced speech mode can fluently speak in multiple languages, and it can even do various accents in English, like a fake French accent.

General AI

Depending on where the conversation is going, I might talk about AGI and ASI.

When talking about the expanding capabilities of computers, you can simply pose the question whether and when will computers be able to accomplish any task that a human can. Say that something like this is called AGI. And if computers become much smarter than the smartest human, that is called ASI. I usually don't go into that.

Another way to illustrate AGI is pointing out that up until recently, AI worked very differently from human intelligence. Human intelligence usually consists of a broad range of core competences, that all or most humans are able to do (learn languages, walk, do the dishes), while each human may specialise in some tasks that they will be good at. You can draw an individual human's competences as a set consisting of this core, and some specialisations. This will look a bit like an amoeba, with a fat core, and some arms going into specialities. For example I say that beyond core competences, I know more than most people about coffee, machine learning research, and a little bit of basketball. Then I draw a different person, who overlaps with me on the core, is much better at basketball, knows nothing about coffee or machine learning research, but is also good at something else I don't know anything about. This is a good opportunity to put in a joke.

Up until recently, AI worked very differently, Each "AI" was a narrow AI with a very limited scope. One AI could be developed to detect pedestrians, a different one to fold proteins, another one to translate from English to Hungarian, another one to play go, etc. You draw this as lots of small bubbles, each representing a different program or model. You can explain that no matter how smart AlphaGo became, and how much better it was at Go than the best humans, it wasn't able to do anything else. Say this is narrow AI.

Current AI tools, especially large language models have changed this picture. These monolithic models, also called foundation models, show remarkable generality. You train a single model, a LLM, and it can than carry out lots of tasks, such as writinig your boring homework, doing your taxes, recommending you movies to watch, explaining you a topic you're interested in, write computer code, whatever.

If you draw language models, they are no longer isolated bubbles, they are a little bit more like human intelligence, in that they, too, share a set of core competences, and this overlaps quite significantly with some of those core competences that we thought were uniquely human. This is why some of these new technologies might feel a little scarier, this is why we are starting to anthropomorphise them.

Last changed by 
 
fhuszar
0
38

Read more

DeepNN Notes on Inductive Biases of Neural Architectures

Lecture notes on general algorithms for calculating Jacobians and gradients, to motivate the use of backpropagation in deep learning.

Feb 6, 2025
DeepNN Notes on The Recent History of Deep Learning

Timeline 1: Breakthroughs in Applications2012 - AlexNet/ImageNet Challenge: AlexNet’s victory in the ImageNet competition marked the resurgence of deep learning, demonstrating the power of CNNs for computer vision.2015 - DQN-Atari: Deep Q-Networks (DQN) achieved human-level performance on Atari games, showcasing the potential of deep reinforcement learning.2015 - Neural Machine Translation (NMT): Google Translate adopted neural machine translation with sequence-to-sequence models and attention, significantly improving translation quality.2015 - Image-to-text captioning: Models like Show, Attend and Tell combined CNNs and RNNs to generate accurate image captions, signaling the rise of multimodal AI.2016 - WaveNet: DeepMind introduced WaveNet, revolutionizing text-to-speech synthesis with natural-sounding, human-like audio.2016 - AlphaGo: DeepMind’s AlphaGo defeated the world champion Go player, demonstrating the potential of reinforcement learning for strategic games.2017 - OpenAI Dota 2 (OpenAI Five): Reinforcement learning models defeated professional players in Dota 2, a complex multiplayer online battle arena game.2019 - AlphaStar (StarCraft II): DeepMind’s AlphaStar reached Grandmaster level in StarCraft II, showcasing AI’s ability to handle real-time strategy games with long time horizons.2019 - GPT-2: OpenAI’s GPT-2 demonstrated the potential of large-scale language models, generating coherent and contextually relevant text.2020 - AlphaFold: DeepMind’s AlphaFold solved the protein folding problem, a decades-old challenge in molecular biology.2021 - AlphaCode/Codex/Copilot (LLMs for text and code): Models like Codex powered applications like GitHub Copilot, revolutionizing programming by enabling natural language coding.2022 - Minerva/LLEMMA/AlphaProof/AlphaGeometry: Advanced models tackled complex mathematical reasoning, formal proofs, and problem-solving in geometry.Timeline 2: MethodsFoundational Architectures (Vision and General Methods):1998 - Basic CNNs: Yann LeCun’s work on convolutional neural networks (CNNs) for handwritten digit recognition (LeNet) established the foundation for modern vision models.2012 - AlexNet: Revived deep learning by scaling up CNNs and using GPUs to win the ImageNet challenge.2014 - VGG: Introduced a very deep architecture with smaller convolutional filters, setting a benchmark for simple, deep CNN designs.2015 - ResNet: Introduced residual connections, enabling the training of very deep networks and surpassing human-level performance in image recognition.2020 - Vision Transformers (ViT): Applied Transformer architectures to vision tasks, achieving state-of-the-art performance by leveraging self-attention mechanisms.Generative Models:2013 - Variational Autoencoders (VAE): Provided a probabilistic framework for generating data while learning latent representations.2014 - Generative Adversarial Networks (GANs): Introduced by Ian Goodfellow, GANs became the foundation for realistic image synthesis and adversarial training.2021 - Diffusion Models: Models like DDPM (Denoising Diffusion Probabilistic Models) enabled state-of-the-art image and audio generation, powering tools like Stable Diffusion and DALL-E 2.Text Models:1997 - LSTM: Long Short-Term Memory networks addressed the vanishing gradient problem, enabling RNNs to capture long-term dependencies.2017 - Attention: The Attention mechanism (introduced in the "Attention is All You Need" paper) allowed models to focus on specific parts of input sequences, revolutionizing sequence modeling.2017 - Transformer: Introduced as a self-attention-based architecture, it became the backbone of nearly all modern NLP and multimodal AI systems.2022 - RLHF (Reinforcement Learning from Human Feedback): Enabled fine-tuning of language models like GPT-3.5 and ChatGPT to align with human intent.2023 - RLAIF (Reinforcement Learning from AI Feedback): Improved alignment efficiency by leveraging model-generated feedback for fine-tuning large language models.Timeline 3: Responsible AI and AI Safety2014 - Adversarial Examples: Szegedy et al. discovered adversarial examples, showing that small, imperceptible changes to input data could drastically alter model predictions, raising concerns about robustness.2016 - Fairness - Equality of Opportunity: The "Equality of Opportunity in Supervised Learning" paper by Hardt et al. introduced algorithmic fairness by ensuring equalized error rates across groups.2016 - Differentially Private SGD: Abadi et al. introduced techniques for training deep learning models with differential privacy, balancing privacy guarantees with utility for sensitive datasets.2016 - GradCAM: Gradient-weighted Class Activation Mapping provided a visual explanation of deep learning models by highlighting important regions in input data.2018 - Gender Shades: Joy Buolamwini and Timnit Gebru’s study revealed biases in commercial facial recognition systems, catalyzing fairness research.2018–2023 - Mechanistic Interpretability: Pioneered by Anthropic’s interpretability research and OpenAI’s circuits work, aiming to reverse-engineer neural network reasoning by analyzing weights and activations.2022 - Constitutional AI: Anthropic introduced Constitutional AI to align language models with human values through rule-based principles, avoiding harmful behaviors without direct human feedback.2023 - Safety Benchmarks (e.g., HELM): The Holistic Evaluation of Language Models formalized evaluation protocols for fairness, bias, and alignment.Timeline 4: Theory of Deep Learning2017 - Understanding Deep Learning Requires Rethinking Generalization: Zhang et al. showed that deep networks can memorize random labels yet still generalize well, challenging traditional theories.2018 - Neural Tangent Kernel (NTK): Jacot et al. showed that deep networks in the infinite-width limit behave like kernel methods, providing insights into gradient descent dynamics.2018 - Deep Linear Models: Saxe et al. studied deep linear networks, simplifying analysis while retaining key properties of deep models.2018–2020 - SGD Implicit Bias: Research showed that stochastic gradient descent (SGD) biases models toward solutions with favorable generalization properties.2019 - Double Descent: Revealed that test error decreases as model capacity grows beyond a critical point, challenging the bias-variance trade-off.2019 - Lottery Ticket Hypothesis: Frankle and Carbin found that large networks contain sparse sub-networks ("winning tickets") that match the performance of the full model.2020 - Scaling Laws: OpenAI and DeepMind demonstrated how model performance scales predictably with size, data, and compute.2021 - Recursive Training: Models can generate high-quality data for self-training, creating a virtuous cycle of performance improvement.2021 - Grokking: Power et al. observed that models sometimes generalize long after fitting the training data, highlighting the interplay of optimization and regularization.2022 - Transcendence (Beyond Memorization): A recent paper explored neural networks’ ability to extrapolate rules beyond training distributions.2022 - Theory of In-context Learning in LLMs: Explored how large language models encode and utilize training distribution statistics to perform implicit meta-learning.2023 - Rule Extrapolation in LLMs: Research formalized how LLMs generalize beyond seen examples, connecting this to abstraction and scaling laws.

Jan 30, 2025
abc conjecture with rationals

My colleague Challenger is working on the abc conjecture. This is one of those problems that are relatively easy to state, but which turn out to be very complicated mathematically.

Dec 12, 2024
# Pytorch basics

# Pytorch basics

Dec 6, 2024
Read more from fhuszar
Published on HackMD